#Quality discrepancy between Dalle-3 and ChatGPT

1 messages · Page 1 of 1 (latest)

frail crane
#

have you guys noticed a huge difference in Dalle-3's ability to generate text onto images when you use ChatGPT vs the api? I'm running the same prompts and the chat gets it right within 1-3 attempts but the api maybe gets it right with 60 attempts
Thats a huge discrepancy for the exact same prompt..

Sample Prompt: Create an image where the theme is: A knight on planet mars make it retro pop. Add the title 'Rumble', clearly visible in text.

royal fiber
frail crane
#

Hey! I was referring to the text prompt I enter NOT the image prompt generated by chatGPT. Being identical is not the problem (im not expecting it to be) but I'm expecting it to spell the word "Rumble" correctly

royal fiber
# frail crane Hey! I was referring to the text prompt I enter NOT the image prompt generated b...

It could be relevant though, I'm speculating. ChatGPT is like an "instruction follower" in its process, where you can say something like "Add the title 'Rumble', clearly visible in text" and ChatGPT will rewrite that in some way for actual image instructions for DALL·E, which is not so much an "instruction follower" in its role as a "mere" text-to-image generator, if that makes sense.

In other words, since DALL·E is a txt2img generator, I think it might be important to test the specific text actually used by DALL·E.

frail crane
#

The problem is the same regardless of which prompt I use. For example if I use my prompt: "make an image of a flying dog with the text 'PAPADAYA'"

or if I use a the generated prompt from ChatGPT Dalle: "A whimsical image of a dog soaring through the sky with its ears and tail fluttering in the wind. The dog has a joyful expression, as it flies among fluffy clouds under a bright blue sky. Bold, cartoon-style letters spell out 'PAPADAYA' in a playful font, floating in the sky alongside the flying dog. The overall feel of the image is light-hearted and magical, capturing a sense of freedom and happiness."

the quality of the image AND the accuracy of the text is remarkable different..

#

There is no question that there is a difference. Regardless of prompt one uses